Introducing Tinker: A Flexible API for Fine-Tuning Language Models

Introducing Tinker: A Flexible API for Fine-Tuning Language Models

October 1, 2025

Today, Thinking Machines is excited to launch Tinker, a new flexible API designed to simplify the process of fine-tuning language models. This tool empowers researchers and developers by providing granular control over algorithms and data while abstracting away the complexities of distributed training. Tinker aligns with our mission to democratize access to cutting-edge AI research and customization.

Key Features and Technical Details

• Model Flexibility: Tinker supports a wide range of models, from small to large, including large mixture-of-experts (MoE) models like Qwen-235B-A22B. Switching between different model sizes is as straightforward as changing a single string in your Python code.
• Managed Service: Running on Thinking Machines' internal clusters and training infrastructure, Tinker handles scheduling, resource allocation, and failure recovery. This allows you to start small or large training runs immediately without worrying about managing the underlying infrastructure.
• Low-Level Primitives: The API provides low-level primitives such as forward_backward and sample, enabling users to implement most common post-training methods. However, achieving optimal results often requires getting many details right, which is why we've also released the Tinker Cookbook-an open-source library with modern implementations of post-training techniques that run on top of the Tinker API.

Use Cases and Early Adopters

Several leading research groups have already started using Tinker:

• Princeton Goedel Team: Trained mathematical theorem provers.
• Rotskoff Chemistry Group at Stanford: Fine-tuned models for chemistry reasoning tasks.
• Berkeley’s SkyRL Group: Conducted experiments on custom async off-policy reinforcement learning (RL) training loops with multi-agents and multi-turn tool use.
• Redwood Research: Utilized Tinker to fine-tune Qwen3-32B on challenging AI control tasks.

Getting Started

Tinker is currently in private beta for researchers and developers. You can sign up for the waitlist here. We will begin onboarding users starting today. If you're an organization interested in using Tinker, please contact us at tinker@thinkingmachines.ai.

Pricing and Availability

Tinker will be free to start. Usage-based pricing will be introduced in the coming weeks.

Technical Implementation Notes

• Distributed Training: Tinker leverages LoRA (Low-Rank Adaptation) to efficiently share compute resources across multiple training runs, reducing costs and improving resource utilization.
• API Design: The API is designed to be intuitive for developers with a basic CS/software background. It includes detailed documentation and examples in the Tinker Cookbook.

Conclusion

We are thrilled to see what the community will achieve with Tinker. Whether you're a researcher exploring new frontiers or a developer customizing models for specific applications, we believe Tinker will provide the tools and flexibility needed to advance your work.